A seismic data denoising method based on time-frequency domain mixed loss constrained convolutional network

Abstract

The present invention discloses a seismic data denoising method of a time-frequency domain mixed loss constrained convolutional network, which relates to the technical field of seismic data processing. The present invention preprocesses seismic data, obtains seismic data samples, constructs a time-frequency domain mixed loss constrained convolutional network model, inputs the seismic data samples into the time-frequency domain mixed loss constrained convolutional network model, uses the seismic data samples to train the time-frequency domain mixed loss constrained convolutional network model to a preset accuracy, obtains the trained time-frequency domain mixed loss constrained convolutional network model, and uses it to process noisy seismic data to obtain the denoising result of the noisy seismic data. The method of the present invention uses a time-frequency domain subnetwork to capture the time domain features and frequency domain features of seismic data in frequency bands, associates the dual-domain features, and uses a dual-domain cross-attention network to dynamically fuse the dual-domain feature information, thereby improving the recognition ability and suppression effect of seismic data noise, and laying the foundation for the accurate interpretation of subsequent seismic data.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic data processing, and in particular to a seismic data denoising method using a time-frequency domain mixed loss-constrained convolutional network. Background Art

[0002] In seismic exploration, raw seismic records are often mixed with a lot of noise (such as environmental noise and instrument noise), and high signal-to-noise ratio data are crucial for subsequent migration imaging and data interpretation.

[0003] Traditional seismic data denoising methods are based on the assumption that the signal and noise are separable in the frequency domain. However, due to the complex spatiotemporal coupling between noise and effective signals in actual data, such methods suffer from problems such as low prediction accuracy and parameter sensitivity. In recent years, deep learning technology, especially convolutional neural networks, has gradually shown its potential in seismic denoising with its powerful feature extraction and long-range dependency modeling capabilities. However, the convolutional network models currently used still have obvious defects. On the one hand, most models only process time domain or frequency domain features separately, failing to fully utilize the complementarity of joint time and frequency information; on the other hand, there is a lack of effective cross-domain feature fusion mechanisms, and no frequency division optimization is designed to address the differences in high- and low-frequency energy distribution in seismic data. All of these lead to serious loss of effective signals.

[0004] Therefore, it is urgent to propose a seismic data denoising method based on a time-frequency domain mixed loss constrained convolutional network, and to improve the seismic data denoising ability of the convolutional network model through joint time-frequency learning and dynamic feature fusion. Summary of the Invention

[0005] The present invention aims to solve the above problems and provides a seismic data denoising method based on a time-frequency domain mixed loss constrained convolutional network. The method adopts a time-frequency domain subnetwork to capture the time domain and frequency domain features of seismic data, establishes a dual-domain feature association, and dynamically fuses the dual-domain feature information through a dual-domain cross-attention network. This achieves efficient identification and suppression of noise in seismic data, improves the seismic data denoising processing capability, and lays the foundation for the accurate interpretation of seismic data.

[0006] The present invention adopts the following technical solutions:

[0007] A seismic data denoising method based on a time-frequency domain mixed loss constrained convolutional network comprises the following steps:

[0008] Step 1: seismic data preprocessing, obtaining seismic data samples;

[0009] Step 2: Construct a time-frequency domain hybrid loss-constrained convolutional network model;

[0010] Step 3: Input the seismic data samples into the time-frequency domain mixed loss constrained convolutional network model, and use the seismic data samples to train the time-frequency domain mixed loss constrained convolutional network model to obtain a trained time-frequency domain mixed loss constrained convolutional network model;

[0011] Step 4: Input the noisy seismic data to be processed into the trained time-frequency domain mixed loss constrained convolutional network model, and use the trained time-frequency domain mixed loss constrained convolutional network model to process the noisy seismic data to obtain the denoising result.

[0012] Preferably, the seismic data preprocessing includes data cutting, data enhancement and data standardization;

[0013] The data cutting adopts a sliding window method to slide and cut the seismic data; the sliding window is a square grid, the sliding step size is consistent with the sliding window width, the seismic data is divided into non-overlapping blocks, and the seismic data is cut along the time axis and the track axis of the seismic data using the sliding window. When the end size of the time axis and the track axis is not enough to cut the seismic data block, the seismic data is rounded up to complete the seismic data;

[0014] The data enhancement is performed by traversing all seismic data blocks, discarding invalid seismic data blocks with all zero values, and performing horizontal flipping along the trace axis on valid seismic data blocks to expand the seismic data;

[0015] The data standardization is performed by traversing all seismic data, determining the global maximum value and the global minimum value in all seismic data, and using the global maximum value and the global minimum value to standardize all seismic data to obtain seismic data samples.

[0016] Preferably, the time-frequency domain mixed loss constrained convolutional network model is provided with a deformable convolution layer, a frequency domain branch network, a time domain sub-network, a dual-domain cross-attention network and an output layer; the deformable convolution layer is respectively connected to the input ends of the frequency domain branch network and the time domain sub-network, and the output ends of the frequency domain branch network and the time domain sub-network are respectively connected to the dual-domain cross-attention network, and the dual-domain cross-attention network is connected to the output layer for outputting the processing results.

[0017] Preferably, the time domain sub-network includes a parallel hole convolution structure, a residual convolution structure and an enhanced residual convolution structure connected in sequence;

[0018] The parallel dilated convolution structure includes four dilated convolution modules with the same structure and a splicing layer, wherein the first dilated convolution module and the second dilated convolution module are set in parallel, the first dilated convolution module and the third dilated convolution module are connected, the second dilated convolution module and the fourth dilated convolution module are connected, and the third dilated convolution module and the fourth dilated convolution module are connected and are all connected to the splicing layer; the dilated convolution module includes a dilated convolution layer, a Relu activation layer and a batch normalization layer connected in sequence, and the processing result of the dilated convolution module is , ,in, For batch normalization, is the two-dimensional standard convolution operation function, is the Relu activation function;

[0019] The residual convolution structure includes two residual convolution modules with the same structure and connected in sequence, and the residual convolution module includes two residual convolution modules connected in sequence. Convolutional layer, Relu activation layer and batch normalization layer;

[0020] The enhanced residual convolution structure includes three enhanced residual convolution modules with the same structure and connected in sequence, and the enhanced residual convolution module includes three enhanced residual convolution modules connected in sequence. Convolutional layer, Relu activation layer and batch normalization layer;

[0021] The frequency domain branch network includes a one-dimensional Fourier transform layer, a binary mask processing layer, a frequency domain sub-network, an inverse Fourier transform layer and a summation layer connected in sequence;

[0022] The frequency domain sub-network has the same structure as the time domain sub-network;

[0023] The dual-domain cross attention network includes a splicing layer, a pooling layer, a one-dimensional convolution layer, Function layer and Convolutional layer.

[0024] Preferably, the seismic data samples are input into a deformable convolution layer, and the deformable convolution layer is used to calculate the offset of the seismic data samples. , determine the offset characteristics of the seismic data sample; the offset ,in, is the offset calculation function, is the earthquake data sample, is a real matrix, is the size of the deformable convolutional layer, is the column of seismic data samples, is the row of seismic data samples;

[0025] Based on the number of output channels of the deformable convolution layer, the sampling points are used along Direction offset position and sampling points along Direction offset position , determine the offset parameters , ,in, is the feature segmentation function; according to the convolution sampling point position Hedi The weight of the convolution kernel calculate , get the initial features , ,in, is the number of output channels;

[0026] The initial features Input into the time domain branch network, first the initial feature Input into the parallel dilated convolution structure, and concatenate the processing results after the first dilated convolution module, the third dilated convolution module, and the processing results after the second dilated convolution module and the fourth dilated convolution module, and then combine them with the initial feature vector. Combined, we get the parallel hole convolution structure processing result , ,in, Represents feature splicing;

[0027] Then process the result of parallel hole convolution structure Input into the residual convolution structure, and combine the processing results of the two residual convolution modules with the processing results of the parallel void convolution structure Add together to get the residual convolution structure processing result ;

[0028] Finally, the residual convolution structure processing results Input into the enhanced residual convolution structure, processed by three enhanced residual convolution modules and output to obtain the time domain sub-network processing result .

[0029] Preferably, the initial features Input into the frequency domain branch network, first the initial features Perform a one-dimensional Fourier transform to obtain a one-dimensional Fourier transform result And obtain its non-negative frequency interval, set three binary masks for frequency division and truncation, and obtain the real and imaginary features of the low frequency band, medium frequency band and high frequency band respectively; the binary mask of the low frequency band is ,in, is the frequency index, is the preset cutoff frequency, which is obtained by element-by-element product of the binary mask of the low-frequency band , ,in, Indicates element-by-element product to extract the real part features of the low-frequency band and imaginary part characteristics , is the column size of the extraction result, , the same method is used to extract the real and imaginary features of the intermediate frequency band and the real and imaginary features of the high frequency band;

[0030] Combine the real and imaginary features of the low-frequency band, medium-frequency band, and high-frequency band respectively to obtain the real and imaginary combined features of each frequency band , ,in, is the real part characteristic, The real and imaginary part features of the low frequency band, medium frequency band and high frequency segment are input into the frequency domain sub-network for processing respectively, and the processing results of each frequency band are inverse Fourier transformed and then summed up to obtain the processing results of the frequency domain branch network. ;

[0031] The time domain sub-network processing results and frequency domain branch network processing results Splicing to obtain time-frequency features , , the time-frequency features Input into the dual-domain cross attention network, the time-frequency features Perform global average pooling operation to obtain the pooling operation result , , Represents the global average pooling operation, which obtains compressed features through dimensional transformation , , and then compress the features Perform one-dimensional convolution and Function operation to obtain attention weight , ,in, is the one-dimensional convolution operation function, for Function; the attention weight and time-frequency characteristics Multiply channel by channel to get the attention result , for attention results Perform point convolution operation to obtain the output result of the time-frequency domain mixed loss constrained convolutional network model .

[0032] Preferably, in step 3, during the training process, each seismic data sample is sequentially input into the time-frequency domain mixed loss constrained convolutional network model, and the time-frequency domain mixed loss constrained convolutional network model is used to predict the low-frequency prediction results, medium-frequency prediction results and high-frequency prediction results of the seismic data sample, and the seismic data sample is decomposed in the frequency domain to obtain the low-frequency results, medium-frequency results and high-frequency results of the seismic data sample, and the total network loss of the time-frequency domain mixed loss constrained convolutional network model is calculated. , ,in, 、 、 、 are weight coefficients, 、 、 、 Both Loss function, using mean absolute error to calculate the loss value, where is the loss value of low-frequency feature output, is the loss value of the intermediate frequency feature output, is the loss value of high-frequency feature output, is the loss value of the time-frequency feature fusion output;

[0033] The time-frequency domain mixed loss constrained convolutional network model is iteratively trained until the total network loss of the time-frequency domain mixed loss constrained convolutional network model converges to a preset accuracy, and the trained time-frequency domain mixed loss constrained convolutional network model is obtained, and the model parameters of the trained time-frequency domain mixed loss constrained convolutional network model are saved.

[0034] The present invention has the following beneficial effects:

[0035] The present invention proposes a seismic data denoising method based on a time-frequency domain mixed loss constrained convolutional network. By constructing a time-frequency domain mixed loss constrained convolutional network model, the time-frequency domain network is used to capture the time-frequency domain characteristics of seismic data and establish dual-domain feature associations. Frequency division optimization is designed based on the difference in high-frequency and low-frequency energy distribution of seismic data. The total network loss is calculated based on the mixed loss to ensure the independence of time-frequency sub-network feature processing. Combined with the dual-domain cross-attention network to dynamically fuse the dual-domain feature information, the noise recognition and noise suppression of seismic data during the processing process are effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flow chart of a seismic data denoising method using a time-frequency domain mixed loss-constrained convolutional network according to the present invention.

[0037] Figure 2 Schematic diagram of the structure of the time-frequency domain mixed loss constrained convolutional network model.

[0038] Figure 3 Schematic diagram of the parallel dilated convolution structure.

[0039] Figure 4 Schematic diagram of the dilated convolution module.

[0040] Figure 5 Schematic diagram of the residual convolution structure.

[0041] Figure 6 Schematic diagram of the enhanced residual convolution structure.

[0042] Figure 7 Schematic diagram of the dual-domain cross-attention network.

[0043] Figure 8 This is a diagram of the denoising effect of the synthetic convolution record; in the figure, (a) is the denoising result of the noisy seismic data, (b) is the pure seismic data, (c) is the waveform of the noisy seismic data after denoising, and (d) is the waveform of the pure seismic data.

[0044] Figure 9 This is the denoising effect diagram of the measured seismic data; in the figure, (a) is the noisy seismic data, (b) is the denoising result of the noisy seismic data, and (c) is the waveform of the seismic data after denoising.

[0045] Figure 10 The figures are comparison diagrams of denoising results of different processing methods; in the figure, (a) is the denoising effect diagram of the general convolutional neural network, (b) is the denoising effect diagram of the FX filtering method, and (c) is the denoising effect diagram of the method of the present invention. DETAILED DESCRIPTION

[0046] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0047] Example 1

[0048] This paper proposes a seismic data denoising method based on a time-frequency domain mixed loss constrained convolutional network. Figure 1 As shown, the following steps are included:

[0049] Step 1: Preprocess seismic data and obtain seismic data samples.

[0050] The seismic data preprocessing includes data cutting, data enhancement and data standardization.

[0051] The data cutting adopts the sliding window method to slide and cut the seismic data; the sliding window is a square grid, the sliding step is consistent with the sliding window width, the seismic data is divided into non-overlapping blocks, and the sliding window is used to slide and cut the seismic data along the time axis and the channel number axis of the seismic data. When the end size of the time axis and the channel number axis is not enough to cut the seismic data block, the seismic data is rounded up to complete the seismic data.

[0052] The data enhancement is performed by traversing all seismic data blocks, discarding invalid seismic data blocks with all zero values, and performing horizontal flipping along the trace axis on valid seismic data blocks to expand the seismic data.

[0053] The data standardization is performed by traversing all seismic data, determining the global maximum value and the global minimum value in all seismic data, and using the global maximum value and the global minimum value to standardize all seismic data to obtain seismic data samples.

[0054] Step 2: Construct a time-frequency domain mixed loss constrained convolutional network model.

[0055] The time-frequency domain mixed loss constrained convolutional network model is as follows Figure 2 As shown, the time-frequency domain mixed loss constrained convolutional network model is provided with a deformable convolution layer, a frequency domain branch network, a time domain sub-network, a dual-domain cross-attention network and an output layer; the deformable convolution layer is respectively connected to the input ends of the frequency domain branch network and the time domain sub-network, and the output ends of the frequency domain branch network and the time domain sub-network are respectively connected to the dual-domain cross-attention network, and the dual-domain cross-attention network is connected to the output layer for outputting the processing results.

[0056] The time domain sub-network includes a parallel hole convolution structure, a residual convolution structure and an enhanced residual convolution structure connected in sequence, wherein the parallel hole convolution structure is as follows: Figure 3 As shown, it includes four dilated convolution modules with the same structure and a splicing layer, wherein the first dilated convolution module and the second dilated convolution module are set in parallel, the first dilated convolution module and the third dilated convolution module are connected, the second dilated convolution module and the fourth dilated convolution module are connected, and the third dilated convolution module and the fourth dilated convolution module are connected and are all connected to the splicing layer; the dilated convolution module is as shown in FIG. Figure 4 As shown in Figure 2, it includes the hole convolution layer, Relu activation layer and batch normalization layer connected in sequence. The processing result of the hole convolution module is , ,in, For batch normalization, is the two-dimensional standard convolution operation function, is the Relu activation function. The residual convolution structure is as follows Figure 5 As shown, it includes two residual convolution modules with the same structure and connected in sequence, and the residual convolution module includes Convolution layer, Relu activation layer and batch normalization layer. The enhanced residual convolution structure is as follows Figure 6 As shown, it includes three enhanced residual convolution modules with the same structure and connected in sequence, and the enhanced residual convolution module includes three enhanced residual convolution modules connected in sequence. Convolutional layer, ReLU activation layer and batch normalization layer.

[0057] The frequency domain branch network includes a one-dimensional Fourier transform layer, a binary mask processing layer, a frequency domain sub-network, an inverse Fourier transform layer and a summation layer connected in sequence, wherein the frequency domain sub-network has the same structure as the time domain sub-network, and the dual-domain cross attention network is as follows Figure 7 As shown, it includes the concatenation layer, pooling layer, one-dimensional convolution layer, Function layer and Convolutional layer.

[0058] Furthermore, when the seismic data sample is input into the deformable convolution layer, the deformable convolution layer is used to calculate the offset of the seismic data sample. , determine the offset characteristics of the seismic data sample; the offset ,in, is the offset calculation function, is the earthquake data sample, is a real matrix, is the size of the deformable convolutional layer, is the column of seismic data samples, is the row of seismic data samples.

[0059] Based on the number of output channels of the deformable convolution layer, the sampling points are used along Direction offset position and sampling points along Direction offset position , determine the offset parameters , ,in, is the feature segmentation function; according to the convolution sampling point position Hedi The weight of the convolution kernel calculate , get the initial features , ,in, is the number of output channels.

[0060] The initial features Input into the time domain branch network, first the initial feature Input into the parallel dilated convolution structure, and concatenate the processing results after the first dilated convolution module, the third dilated convolution module, and the processing results after the second dilated convolution module and the fourth dilated convolution module, and then combine them with the initial feature vector. Combined, we get the parallel hole convolution structure processing result , ,in, Represents feature splicing;

[0061] Then process the result of parallel hole convolution structure Input into the residual convolution structure, and combine the processing results of the two residual convolution modules with the processing results of the parallel void convolution structure Add together to get the residual convolution structure processing result ;

[0062] Finally, the residual convolution structure processing results Input into the enhanced residual convolution structure, processed by three enhanced residual convolution modules and output to obtain the time domain sub-network processing result .

[0063] At the same time, the initial features Input into the frequency domain branch network, first the initial features Perform a one-dimensional Fourier transform to obtain a one-dimensional Fourier transform result And obtain its non-negative frequency interval, set three binary masks for frequency division and truncation, and obtain the real and imaginary features of the low frequency band, medium frequency band and high frequency band respectively.

[0064] In this embodiment, the low frequency band is 0-25 Hz, the medium frequency band is 25 Hz-40 Hz, and the high frequency band is above 40 Hz; the binary mask of the low frequency band is ,in, is the frequency index, is the preset cutoff frequency, which is obtained by element-by-element product of the binary mask of the low-frequency band , ,in, Indicates element-by-element product to extract the real part features of the low-frequency band and imaginary part characteristics , is the column size of the extraction result, The same method is used to extract the real and imaginary features of the intermediate frequency band and the real and imaginary features of the high frequency band.

[0065] Then, the real and imaginary features of the low-frequency band, medium-frequency band, and high-frequency band are combined to obtain the real and imaginary combined features of each frequency band. , ,in, is the real part characteristic, The real and imaginary part features of the low frequency band, medium frequency band and high frequency segment are input into the frequency domain sub-network for processing respectively, and the processing results of each frequency band are inverse Fourier transformed and then summed up to obtain the processing results of the frequency domain branch network. .

[0066] Finally, the time domain sub-network processing results and frequency domain branch network processing results Splicing to obtain time-frequency features , , the time-frequency features Input into the dual-domain cross attention network, the time-frequency features Perform global average pooling operation to obtain the pooling operation result , , Represents the global average pooling operation, which obtains compressed features through dimensional transformation , , and then compress the features Perform one-dimensional convolution and Function operation to obtain attention weight , ,in, is the one-dimensional convolution operation function, for Function; the attention weight and time-frequency characteristics Multiply channel by channel to get the attention result , for attention results Perform point convolution operation to obtain the output result of the time-frequency domain mixed loss constrained convolutional network model .

[0067] Step 3: Input the seismic data samples into the time-frequency domain mixed loss constrained convolutional network model, use the seismic data samples to train the time-frequency domain mixed loss constrained convolutional network model, and obtain the trained time-frequency domain mixed loss constrained convolutional network model.

[0068] During the training process, each earthquake data sample is input into the time-frequency domain mixed loss constrained convolutional network model in turn, and the low-frequency prediction results of the earthquake data sample are obtained by using the time-frequency domain mixed loss constrained convolutional network model. , medium frequency prediction results and high-frequency prediction results , and perform frequency domain decomposition on the seismic data samples to obtain the low-frequency results of the seismic data samples , intermediate frequency results and high-frequency results , calculate the total network loss of the time-frequency domain mixed loss constrained convolutional network model , ,in, 、 、 、 are weight coefficients, 、 、 、 Both Loss function, using mean absolute error to calculate the loss value, where is the loss value of low-frequency feature output, is the loss value of the intermediate frequency feature output, is the loss value of high-frequency feature output, is the loss value of the time-frequency feature fusion output.

[0069] In this embodiment, the mean absolute error is calculated for the low-frequency band, the medium-frequency band, and the high-frequency band respectively, and the time-frequency domain mixed loss constrained convolutional network model is iteratively trained until the total network loss of the time-frequency domain mixed loss constrained convolutional network model converges to a preset accuracy, and the trained time-frequency domain mixed loss constrained convolutional network model is obtained, and the model parameters of the trained time-frequency domain mixed loss constrained convolutional network model are saved.

[0070] Step 4: Input the noisy seismic data to be processed into the trained time-frequency domain mixed loss constrained convolutional network model, and use the trained time-frequency domain mixed loss constrained convolutional network model to process the noisy seismic data to obtain the denoising result.

[0071] Example 2

[0072] In order to verify the denoising effect of the method of the present invention, the seismic data denoising method of the time-frequency domain mixed loss constrained convolutional network described in Example 1 was used to process actual seismic data.

[0073] In this embodiment, when preprocessing seismic data, a sliding window method is used to divide the continuous seismic traces in the seismic data into seismic data of size 256×256, and the seismic data samples are horizontally flipped and expanded along the trace axis. Finally, the amplitude value is scaled to the interval [0, 1] before inputting the time-frequency domain mixed loss constrained convolutional network model. The preprocessed seismic data samples are input into the time-frequency domain mixed loss constrained convolutional network model, and the real and imaginary features of the preset low-frequency band, medium-frequency band, and high-frequency band are intercepted. The time-frequency domain mixed loss constrained convolutional network model is trained using the preprocessed seismic data samples, and the time-frequency domain mixed loss constrained convolutional network model is converged according to whether the total network loss of the time-frequency domain mixed loss constrained convolutional network model reaches the preset accuracy, and the trained time-frequency domain mixed loss constrained convolutional network model is obtained. The trained time-frequency domain mixed loss constrained convolutional network model is used to denoise the actual noisy seismic data, and the following is obtained: Figure 8 The denoising effect of the synthetic convolution record is shown. From the denoising effect of the synthetic convolution record, it can be seen that the method of the present invention can effectively achieve seismic data noise suppression and signal reconstruction. Figure 9The denoising effect of the method of the present invention on measured seismic data is shown in Figure 2. By comparison, it can be found that the method of the present invention can suppress background noise while restoring details in the seismic signal as much as possible. In order to further verify the superiority of the denoising effect of the method of the present invention, the denoising effect of the method of the present invention is compared with that of the general convolutional neural network and FX filtering methods. The same noisy seismic data are processed and compared using the method of the present invention, the general convolutional neural network and the FX filtering method. The comparison results are shown in Figure 2. Figure 10 As shown in Table 1, the denoising evaluation indicators of each processing method are compared. FX filtering method under different parameter conditions.

[0074] Table 1 Denoising evaluation index table

[0075] Denoising evaluation indicators Noisy data FX(1) FX(2) FX(3) This method model Peak signal-to-noise ratio 17.238 19.249 20.292 20.922 24.385 Signal-to-noise ratio -0.043 1.967 3.010 3.641 7.103

[0076] After comparing the peak signal-to-noise ratio and signal-to-noise ratio of different methods, it was found that when a general convolutional neural network is used for seismic data denoising, signal attenuation and event continuity destruction will occur in the strong amplitude region of the seismic data, while the time-frequency domain mixed loss-constrained convolutional network model used in the method of the present invention has better signal retention capability, and the method of the present invention can effectively improve the signal-to-noise ratio and peak signal-to-noise ratio of seismic data when applied to seismic data denoising.

[0077] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A seismic data denoising method based on a time-frequency domain mixed loss constrained convolutional network, characterized in that: The following steps are involved: Step 1: seismic data preprocessing, obtaining seismic data samples; Step 2: Construct a time-frequency domain hybrid loss-constrained convolutional network model; Step 3: Input the seismic data samples into the time-frequency domain mixed loss constrained convolutional network model, and use the seismic data samples to train the time-frequency domain mixed loss constrained convolutional network model to obtain a trained time-frequency domain mixed loss constrained convolutional network model; Step 4: input the noisy seismic data to be processed into the trained time-frequency domain mixed loss constrained convolutional network model, and use the trained time-frequency domain mixed loss constrained convolutional network model to process the noisy seismic data to obtain a denoising result; The time-frequency domain mixed loss constrained convolutional network model is provided with a deformable convolution layer, a frequency domain branch network, a time domain sub-network, a dual-domain cross attention network and an output layer; the deformable convolution layer is respectively connected to the input ends of the frequency domain branch network and the time domain sub-network, the output ends of the frequency domain branch network and the time domain sub-network are respectively connected to the dual-domain cross attention network, and the dual-domain cross attention network is connected to the output layer for outputting the processing results; The time domain sub-network includes a parallel hole convolution structure, a residual convolution structure and an enhanced residual convolution structure connected in sequence; Input the seismic data sample into the deformable convolution layer, and use the deformable convolution layer to calculate the offset of the seismic data sample , determine the offset characteristics of the seismic data sample; Based on the number of output channels of the deformable convolution layer, the sampling points are used along Direction offset position and sampling points along Direction offset position , determine the offset parameters , ,in, is the feature segmentation function; according to the convolution sampling point position Hedi The weight of the convolution kernel calculate , get the initial features , ,in, is the number of output channels; The initial features Input into the time domain sub-network to obtain the time domain sub-network processing results ; The initial features Input into the frequency domain branch network, first the initial features Perform a one-dimensional Fourier transform to obtain a one-dimensional Fourier transform result And obtain its non-negative frequency interval, set three binary masks for frequency division and truncation, and obtain the real and imaginary features of the low frequency band, medium frequency band and high frequency band respectively; Combine the real and imaginary features of the low frequency band, medium frequency band, and high frequency band respectively to obtain the real and imaginary combined features of each frequency band , ,in, is the real part characteristic, The real and imaginary part features of the low frequency band, the medium frequency band, and the high frequency band are respectively input into the frequency domain branch network for processing, and the processing results of each frequency band are inverse Fourier transformed and then summed up to obtain the frequency domain branch network processing results. .

2. The seismic data denoising method using a time-frequency domain mixed loss-constrained convolutional network according to claim 1 is characterized in that: The seismic data preprocessing includes data cutting, data enhancement and data standardization; The data cutting adopts a sliding window method to slide and cut the seismic data; the sliding window is a square grid, the sliding step size is consistent with the sliding window width, the seismic data is divided into non-overlapping blocks, and the seismic data is cut along the time axis and the track axis of the seismic data using the sliding window. When the end size of the time axis and the track axis is not enough to cut the seismic data block, the seismic data is rounded up to complete the seismic data; The data enhancement is performed by traversing all seismic data blocks, discarding invalid seismic data blocks with all zero values, and performing horizontal flipping along the trace axis on valid seismic data blocks to expand the seismic data; The data standardization is performed by traversing all seismic data, determining the global maximum value and the global minimum value in all seismic data, and using the global maximum value and the global minimum value to standardize all seismic data to obtain seismic data samples.

3. The seismic data denoising method using a time-frequency domain mixed loss-constrained convolutional network according to claim 1 is characterized in that: The parallel dilated convolution structure includes four dilated convolution modules with the same structure and a splicing layer, wherein the first dilated convolution module and the second dilated convolution module are set in parallel, the first dilated convolution module and the third dilated convolution module are connected, the second dilated convolution module and the fourth dilated convolution module are connected, and the third dilated convolution module and the fourth dilated convolution module are connected and are all connected to the splicing layer; the dilated convolution module includes a dilated convolution layer, a Relu activation layer and a batch normalization layer connected in sequence, and the processing result of the dilated convolution module is , ,in, For batch normalization, is the two-dimensional standard convolution operation function, is the Relu activation function; The residual convolution structure includes two residual convolution modules with the same structure and connected in sequence, and the residual convolution module includes two residual convolution modules connected in sequence. Convolutional layer, Relu activation layer and batch normalization layer; The enhanced residual convolution structure includes three enhanced residual convolution modules with the same structure and connected in sequence, and the enhanced residual convolution module includes three enhanced residual convolution modules connected in sequence. Convolutional layer, Relu activation layer and batch normalization layer; The frequency domain branch network includes a one-dimensional Fourier transform layer, a binary mask processing layer, a frequency domain sub-network, an inverse Fourier transform layer and a summation layer connected in sequence; The frequency domain sub-network has the same structure as the time domain sub-network; The dual-domain cross attention network includes a splicing layer, a pooling layer, a one-dimensional convolution layer, Function layer and Convolutional layer.

4. The seismic data denoising method using a time-frequency domain mixed loss-constrained convolutional network according to claim 3, characterized in that: The offset ,in, is the offset calculation function, is the earthquake data sample, is a real matrix, is the size of the deformable convolutional layer, is the column of seismic data samples, is the row of seismic data samples; The initial features Input into the time domain sub-network, first the initial features Input into the parallel dilated convolution structure, and concatenate the processing results after the first dilated convolution module, the third dilated convolution module, and the processing results after the second dilated convolution module and the fourth dilated convolution module, and then combine them with the initial feature vector. Combined, we get the parallel hole convolution structure processing result , ,in, Represents feature splicing; Then process the result of parallel hole convolution structure Input into the residual convolution structure, and combine the processing results of the two residual convolution modules with the processing results of the parallel void convolution structure Add together to get the residual convolution structure processing result ; Finally, the residual convolution structure processing results Input into the enhanced residual convolution structure, processed by three enhanced residual convolution modules and output to obtain the time domain sub-network processing result .

5. The seismic data denoising method using a time-frequency domain mixed loss-constrained convolutional network according to claim 4 is characterized in that: The binary mask of the low frequency band is ,in, is the frequency index, is the preset cutoff frequency, which is obtained by element-by-element product of the binary mask of the low-frequency band , ,in, Indicates element-by-element product, extracting the real part features of the low-frequency band and imaginary part characteristics , is the column size of the extraction result, , the same method is used to extract the real and imaginary features of the intermediate frequency band and the real and imaginary features of the high frequency band; The time domain sub-network processing results and frequency domain branch network processing results Splicing to obtain time-frequency features , , the time-frequency features Input into the dual-domain cross attention network, the time-frequency features Perform global average pooling operation to obtain the pooling operation result , , Represents the global average pooling operation, which obtains compressed features through dimensional transformation , , and then compress the features Perform one-dimensional convolution and Function operation to obtain attention weight , ,in, is the one-dimensional convolution operation function, for Function; the attention weight and time-frequency characteristics Multiply channel by channel to get the attention result , for attention results Perform point convolution operation to obtain the output result of the time-frequency domain mixed loss constrained convolutional network model .

6. The seismic data denoising method using a time-frequency domain mixed loss-constrained convolutional network according to claim 1 is characterized in that: In step 3, during the training process, each seismic data sample is sequentially input into the time-frequency domain mixed loss constrained convolutional network model, and the time-frequency domain mixed loss constrained convolutional network model is used to predict the low-frequency prediction results, medium-frequency prediction results and high-frequency prediction results of the seismic data sample, and the seismic data sample is decomposed in the frequency domain to obtain the low-frequency results, medium-frequency results and high-frequency results of the seismic data sample, and the total network loss of the time-frequency domain mixed loss constrained convolutional network model is calculated. , ,in, 、 、 、 are weight coefficients, 、 、 、 Both Loss function, using mean absolute error to calculate the loss value, where is the loss value of low-frequency feature output, is the loss value of the intermediate frequency feature output, is the loss value of high-frequency feature output, is the loss value of the time-frequency feature fusion output; The time-frequency domain mixed loss constrained convolutional network model is iteratively trained until the total network loss of the time-frequency domain mixed loss constrained convolutional network model converges to a preset accuracy, and the trained time-frequency domain mixed loss constrained convolutional network model is obtained, and the model parameters of the trained time-frequency domain mixed loss constrained convolutional network model are saved.

Images